Re: [PATCH] EH_USES df fix (PR target/37378, take 2)

[Kenneth Zadeck <zadeck at naturalbridge dot com>]

Jakub Jelinek wrote:
> Hi!
>
> As pointed out by Hans-Peter Nilsson, we want to make EH_USES regs live even
> if some code path ends with an endless loop, the insns in between still
> might throw, or during debugging the debugger (or something called from
> async signal handler) might want to do a backtrace).  The patch I've posted
> earlier doesn't cure e.g.
>
>   
i seriously doubt that this patch will bootstrap on a non ia-64 target,
but once you fix that, the patch is ok.   (the problem is that the
declaration of i needs to be in an ifdef EH_USES.)


> extern int foo (int, int);
> extern int bar (void);
>
> int
> baz (int x, int y)
> {
>   if (__builtin_expect (y == 0, 0))
>     {
>       foo (x + y, x * y);
>       bar ();
>       while (1)
> 	;
>     }
>
>   if (x == (int) 0x80000000L && y == -1)
>     return x;
>
>   return x / y;
> }
> at -O2 - rp is saved into a local register only if y != 0.
> The following alternative patch makes EH_USES registers just live
> everywhere, which IMHO matches best what we really want - let those
> registers stay live through the whole function.
>
> Bootstrapped/regtested on ia64-linux, ok for trunk?
>
> 2008-10-27  Jakub Jelinek  <jakub@redhat.com>
>
> 	PR target/37378
> 	* df-scan.c (df_bb_refs_collect): Don't handle EH_USES here.
> 	(df_get_entry_block_def_set): Neither here.
> 	(df_get_regular_block_artificial_uses): Add EH_USES registers.
>
> --- gcc/df-scan.c.jj	2008-10-14 07:58:50.000000000 -0400
> +++ gcc/df-scan.c	2008-10-27 14:13:41.000000000 -0400
> @@ -3555,29 +3555,6 @@ df_bb_refs_collect (struct df_collection
>      }
>  #endif
>  
> -
> -#ifdef EH_USES
> -  if (bb_has_eh_pred (bb))
> -    {
> -      unsigned int i;
> -      /* This code is putting in an artificial ref for the use at the
> -	 TOP of the block that receives the exception.  It is too
> -	 cumbersome to actually put the ref on the edge.  We could
> -	 either model this at the top of the receiver block or the
> -	 bottom of the sender block.
> -
> -         The bottom of the sender block is problematic because not all
> -         out-edges of a block are eh-edges.  However, it is true
> -         that all edges into a block are either eh-edges or none of
> -         them are eh-edges.  Thus, we can model this at the top of the
> -         eh-receiver for all of the edges at once. */
> -      for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
> -	if (EH_USES (i))
> -	  df_ref_record (DF_REF_ARTIFICIAL, collection_rec, regno_reg_rtx[i], NULL,
> -			 bb, NULL, DF_REF_REG_USE, DF_REF_AT_TOP, -1, -1, 0);
> -    }
> -#endif
> -
>    /* Add the hard_frame_pointer if this block is the target of a
>       non-local goto.  */
>    if (bb->flags & BB_NON_LOCAL_GOTO_TARGET)
> @@ -3667,6 +3644,8 @@ df_bb_refs_record (int bb_index, bool sc
>  static void
>  df_get_regular_block_artificial_uses (bitmap regular_block_artificial_uses)
>  {
> +  unsigned int i;
> +
>    bitmap_clear (regular_block_artificial_uses);
>  
>    if (reload_completed)
> @@ -3702,6 +3681,20 @@ df_get_regular_block_artificial_uses (bi
>      }
>    /* The all-important stack pointer must always be live.  */
>    bitmap_set_bit (regular_block_artificial_uses, STACK_POINTER_REGNUM);
> +
> +#ifdef EH_USES
> +  /* EH_USES registers are used:
> +     1) at all insns that might throw (calls or with -fnon-call-exceptions
> +	trapping insns)
> +     2) in all EH edges
> +     3) to support backtraces and/or debugging, anywhere between their
> +	initialization and where they the saved registers are restored
> +	from them, including the cases where we don't reach the epilogue
> +	(noreturn call or infinite loop).  */
> +  for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
> +    if (EH_USES (i))
> +      bitmap_set_bit (regular_block_artificial_uses, i);
> +#endif
>  }
>  
>  
> @@ -3826,16 +3819,6 @@ df_get_entry_block_def_set (bitmap entry
>    /* These registers are live everywhere.  */
>    if (!reload_completed)
>      {
> -#ifdef EH_USES
> -      /* The ia-64, the only machine that uses this, does not define these 
> -	 until after reload.  */
> -      for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
> -	if (EH_USES (i))
> -	  {
> -	    bitmap_set_bit (entry_block_defs, i);
> -	  }
> -#endif
> -      
>  #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM
>        /* Pseudos with argument area equivalences may require
>  	 reloading via the argument pointer.  */
>
>
> 	Jakub
>